'Horizon Forbidden West' GDC session covers studio's switch to open-source physics engine

cormack12

Gold Member
Source: https://www.gamedeveloper.com/gdc/-...studio-s-switch-to-open-source-physics-engine

Looks like Gueriila are going to be doing a lot this year again. Really enjoyed their Killzone talks (https://schedule.gdconf.com/search/horizon+forbidden+west/ )

Guerrilla Games is back with Horizon Forbidden West, the highly anticipated follow-up to its 2017 video game about fighting robot dinosaurs in a post-apocalypse future. At GDC 2022, the studio will be covering everything from character faces to storms and bunkers—as well as the studio's exciting switch to an open-source physics engine.

During this session, lead game tech Jorrit Rouwe will describe Guerrilla Games' switch from a commercial physics engine to Jolt Physics, an open-source physics engine that was born as a personal hobby project. By switching to this new engine, the studio saved memory, executable size, and were able to double their simulation frequency while using less CPU time.

Jorrit will cover how they use a physics engine at Guerrilla, how their previous physics engine caused bottlenecks while streaming in data and while interacting with the multi-thread game object update. Jolt was architected specifically to help solve those two problems, so be sure to check out this session to learn about two of the systems that made this possible.

GDC returns in-person to San Francisco, March 21-25, 2022—registration is now open!
 

DeepEnigma

Gold Member
Why not use something like havok or physx? I mean they have years of experience and battle testing
During this session, lead game tech Jorrit Rouwe will describe Guerrilla Games' switch from a commercial physics engine to Jolt Physics, an open-source physics engine that was born as a personal hobby project. By switching to this new engine, the studio saved memory, executable size, and were able to double their simulation frequency while using less CPU time.

Jorrit will cover how they use a physics engine at Guerrilla, how their previous physics engine caused bottlenecks while streaming in data and while interacting with the multi-thread game object update.
HZD used Havok.
 
Last edited:

Kuranghi

Gold Member
I do like the effect when you pull down walls and such, looks nice. I'm really hoping for physics to start having better transitions from static geometry to the physics sim and then to a sleeping physics state and back to being active possibly as well.

The hair is a bit crazy sometimes, intersecting when shes bending over, like the cutscene where she puts her head in her hands near the start but I really like it for the most part.
 

GymWolf

Gold Member
Not impressed by the physics\destruction very much unfortunately, rocks break a bit more realistically compared to the first game i guess...maybe there is something more advanced going forward with the story but i doubt it.

I expect a major overhaul in the third game when they are not gonna be fucked in the ass by a jaguar anymore.
 

CamHostage

Member
I do like the effect when you pull down walls and such, looks nice. I'm really hoping for physics to start having better transitions from static geometry to the physics sim and then to a sleeping physics state and back to being active possibly as well.

Oh yeah, huh? Physics, that was a big thing of "not next-gen enough" opinions in the initial showcase, where the physics engine was caught eliminating fragments after they were done animating. I haven't heard anybody talk about the physics system since (and we've not had any cases of so-called "next-gen" physics usage, especially in a commercial game of a grand scale like this, where some new level of object persistence was demonstrated never possible before... it happens in everything if you have to balance all the other things going on along with the physics, even in the upcoming UE5 and its implementation of Chaos physics system, depending on how you use it.)

I'm curious how people are finding the final game to have worked out with the physics (don't have a PS5 to play it on yet,) obviously this wasn't such an issue that people are still bugging out about it but how does their implementation compare over the showcase version or over H:ZD? (I'm watching a video now and it seems pretty much the same as the first game in how it handles debris unfortunately, but oh well.)

 
Last edited:

STARSBarry

Gold Member
Could this be why Aloys jumping etc is so shit? I have mentioned how much the new climbing is bugging me largely due to the leaps and jumps where it misses an obvious grapple point. If it's linked to this physics system could explain why it's so awful.
 

CamHostage

Member
BTW, I don't think this'll be very exciting to check out as it's a technical demo reel rather than a "showcase" showing off amazing things the physics system could do, but there is a playlist from 2021 showing some different base implementations of Jolt Physics. Take a look through the reel if you're technically-minded and can see something the rest of us might not understand...


I expect a major overhaul in the third game when they are not gonna be fucked in the ass by a jaguar anymore.

I mean, it'd be nice and I hope something comes along that demonstrates a new level of persistent physics efficiency and complexity, but I really think people are overblowing the idea that cross-gen is the thing holding everything back, that there will be an amazing outpouring of unimaginable effects quality once more games go next-gen-only. Returnal is next-gen only and its fractured objects fragments still disappear; objects in the upcoming next-gen only Hellblade 2 also disappear once their baked physics animation is completed (it's hard to see but in the Hellblade 2 Gameplay Reveal you can catch objects disappearing after they bounce on the ground, with only hero-particles like the rolling log chunk staying resident.) Depending on what type of game and how its used, Unreal Chaos Physics destruction will do the same thing. Something's always got to give. I feel like Horizon FW could have done better (disappearing right in front of your face is always a big no-no) but when a game is made on such a scale, physical destruction is just one thing in a massive mix.

 
Last edited:

GymWolf

Gold Member
BTW, I don't think this'll be very exciting to check out as it's a technical demo reel rather than a "showcase" showing off amazing things the physics system could do, but there is a playlist from 2021 showing some different base implementations of Jolt Physics. Take a look through the reel if you're technically-minded and can see something the rest of us might not understand...




I mean, it'd be nice and I hope something comes along that demonstrates a new level of persistent physics efficiency and complexity, but I really think people are overblowing the idea that cross-gen is the thing holding everything back, that there will be an amazing outpouring of unimaginable effects quality once more games go next-gen-only. Returnal is next-gen only and its fractured objects fragments still disappear; objects in the upcoming next-gen only Hellblade 2 also disappear once their baked physics animation is completed (it's hard to see but in the Hellblade 2 Gameplay Reveal you can catch objects disappearing after they bounce on the ground, with only hero-particles like the rolling log chunk staying resident.) Depending on what type of game and how its used, Unreal Chaos Physics destruction will do the same thing. Something's always got to give.

Of course, but physics\destruction in h2 is 99% identical to horizon 1, i really think that a modern ryzen can do much more than a jaguar.
 
Why not use something like havok or physx? I mean they have years of experience and battle testing
Why use the Havok ore Physx engine? They must pay a license feed for every time they use it. They build theyr own engine, they have the the expertise in house to build theyr own ( open source) physics engine and they did. And its open source, thats more flexibele than a close source engine like Havok ore Physx....
 
Of course, but physics\destruction in h2 is 99% identical to horizon 1, i really think that a modern ryzen can do much more than a jaguar.

Isn't the game required to run on those same Jaguar's?

It sucks, but give it time. Good things to come. Really looking forward to the talk though. Thanks for sharing!
 
Last edited:

GymWolf

Gold Member
Isn't the game required to run on those same Jaguar's?

It sucks, but give it time. Good things to come. Really looking forward to the talk though. Thanks for sharing!
Yes, that's why i said that the idea of crossgen slowing down physics in games is not overblown, h2 has the same physics of H1 because the game has to run on a jaguar.

I expect big things from native nextgen games on both consoles.
 

Three

Member
Why not use something like havok or physx? I mean they have years of experience and battle testing
Havok was bought by MS during the xbox one gen and Physx is Nvidia proprietary. Jolt probably made the most sense business wise.
 
Not impressed by the physics\destruction very much unfortunately, rocks break a bit more realistically compared to the first game i guess...maybe there is something more advanced going forward with the story but i doubt it.

I expect a major overhaul in the third game when they are not gonna be fucked in the ass by a jaguar anymore.

But if it's true that this engine is much lighter than what they where using before, it explain how they could extract even more from PS4.
 

GymWolf

Gold Member
But if it's true that this engine is much lighter than what they where using before, it explain how they could extract even more from PS4.
I guess that they had some spare power to upgrade the graphic but physics\destruction really isn't that much upgraded except vegetation reacting to aloy passage.
 

GymWolf

Gold Member
The trees seem to sway with the wind more far this time around. Same with foliage.
Aren't those just precanned patterns? vfx veteran explained to me the same tecnique used in tsushima.

It's not really dynamic physics, it is an effect that simulate that, i think it weight more on the gpu than cpu.

Also, trees moving with the wind is not a new thing or anything, many games have this feature.
 

SlimySnake

The Contrarian
Also, trees moving with the wind is not a new thing or anything, many games have this feature.
Yeah, but just like with bad water, their engine did not support features we have seen in other games.

As for a CPU vs GPU, i think you need both for physics. I used to run Nvidia Phsysx in Arkham and Ass Creed games on PC last gen. Would take a massive hit in performance.
 

yurinka

Member
Why not use something like havok or physx? I mean they have years of experience and battle testing
Because they did use Havok and as you can see reading the description of the talk, they found a better engine. Physx is only for Nvidia.
 
Last edited:

Dr Bass

Member
Aren't those just precanned patterns? vfx veteran explained to me the same tecnique used in tsushima.

It's not really dynamic physics, it is an effect that simulate that, i think it weight more on the gpu than cpu.

Also, trees moving with the wind is not a new thing or anything, many games have this feature.
It has to surely be pre canned patterns. I cant imagine there is any way in hell a physics solver is running on every tree/branch/leaf.
 

CamHostage

Member
Yes, that's why i said that the idea of crossgen slowing down physics in games is not overblown, h2 has the same physics of H1 because the game has to run on a jaguar.

Horizon FW has the same general physics as Horizon ZD because it's using the same techniques; that's their house style on their own engine. That's what they're used to working with, with improvements across the board in density and complexity both in scaling from PS4 to PS5 as well as general improvements going into a second game, but it's a minmax of what they know how to do and how to improve doing this all a second time. I don't think it has much to do with it having to run on Jaguar CPU, they had overhead room to scale it here (and they did in the vegetation) and this still is what they got out of it.

Getting different results would mean an overhaul of the physics in general, (which yes please, but) that takes time and development, not simply horsepower. Some techniques may take advantage of the significantly faster processor and compute units and whatnot (tech is my interest but not strong suit, perhaps obviously) to do things never possible before, but that's a rare breakthrough these days; more often, it's just cramming more passes and more layers and more elegant or ML-honed methods into the flow of the hardware (sometimes still in ways that the past-gen hardware can take advantage of once it's been worked out on a larger scale,) and larger hardware like this generation of consoles gives a greater pipe for that flow.

Guerrilla don't have that right now (albeit maybe that's in the lab for future projects on Decima,) and they didn't have a game which required a rewrite of all of that; they simply needed more of and refinements over what they already knew how to do. We have our Nubis weather system for making volumetric clouds, great; now let's see if we can push that to have multiple layers and fog which rolls off of hillsides and low-laying clouds. We have our water surface system (with improvements added to Frozen Wilds) for showing lakes and rivers, great; now let's push it and make oceans and have areas under the surface and have ocean waves billowing up seafoam. We have our landscape system and procedural vegetation methods depicting large areas flocked with plants and trees, great; let's push that to the point where we can glide and fly over an area and can still maintain the integrity of the world from sky to ground. What Decima does is good, industry-leading even, and it's foundational to both cross-gen and future next-gen projects. So they made improvements, not radical reinventions, and while there are certainly things they could have done which may have been better if they had pushed the physics engine (either cross-gen or exclusively for a PS5 version,) that wasn't in the scope of this project (especially given its original target of somewhere around Q3 2021.) Maybe a future Guerrilla or other Decima game will introduce radical reinventions, but most likely Decima is their past and their present and their future, and there is no Decima 2 around the corner because there's still wide room to grow with Decima as they make advancements/improvements to it.

I expect big things (in physics) from native nextgen games on both consoles.

I do too (and it has been frustrating waiting for games that demonstrate that,) but that'll be coming from either significant pushes on pre-existing tech in new projects that overhaul the way it works or innovate in use of those ways (and probably those projects will all be smaller in scale than H:FW) or new techniques (probably trained in ML simulation) which are still in the labs. Unlike some past generations of consoles, there wasn't any super-advanced physics system (not that I'm aware of at least) which was just waiting for next-gen hardware to be released so it could be included in brand new games. There should be games which push physics in ways not seen on previous consoles (I recommend watching Liquid Crystal, a project by Grant Kot, a current/former(?) GAFer, built around a liquid physics and destructible MagicaVoxel structures, with a target of Xbox Series X-level hardware,) but that'll be because they put their focus on improving and showcasing that aspect of their game.

More horsepower doesn't necessarily mean you're ready to reinvent the wheel.
 
Last edited:

Three

Member
Aren't those just precanned patterns? vfx veteran explained to me the same tecnique used in tsushima.

It's not really dynamic physics, it is an effect that simulate that, i think it weight more on the gpu than cpu.

Also, trees moving with the wind is not a new thing or anything, many games have this feature.
What did VFXVeteran tell you? He's probably wrong. They've used physics for their wind animations since HZD.

You can learn about it here
 

Truespeed

Member
No brainer. No licensing fee and Jolt seems to scale better than its commercial competitors. Also, it's open source and available on GitHub. Don't like how a certain thing works? Fork it and customize it.
 

Fafalada

Fafracer forever
Aren't those just precanned patterns?
There's no one method to do any of this. But most approaches you see in modern games involve some form of force-integration and collision intersection tests - which is well - physics.

I do too (and it has been frustrating waiting for games that demonstrate that,) but that'll be coming from either significant pushes on pre-existing tech in new projects that overhaul the way it works or innovate in use of those ways (and probably those projects will all be smaller in scale than H:FW) or new techniques (probably trained in ML simulation) which are still in the labs.
I mean - we've been riding that wave for over 2 decades now - with physics being done mostly the same way, with similar level of interactivity, since PS2 era.
Partly the problem is that simulation workloads scale exponentially - rather than linearly like the rendering does, and partly, game design just haven't come up with meaningful advancements on the back of more dynamic game worlds (beyond making presentation prettier - which is really just another way of improving visuals...).
I do think ML opens up some interesting new possibilities for simulation scaling - but again, majority of those use cases would be eye-candy rather than anything more impactful.
 
Last edited:
We’re seeing realistic physics done best with voxel based engines, Teardown is a good example and next gen for sure. Doubt a console could run it well.
 

VFXVeteran

Banned
I do like the effect when you pull down walls and such, looks nice. I'm really hoping for physics to start having better transitions from static geometry to the physics sim and then to a sleeping physics state and back to being active possibly as well.

The hair is a bit crazy sometimes, intersecting when shes bending over, like the cutscene where she puts her head in her hands near the start but I really like it for the most part.
That is beyond realtime hardware. You are talking about completely bounding every single strand of hair (or hair clump) and testing for collisions. The more accurate you want it to be the more intersection tests with small bounds will be needed. After that what do you do when you have to compute it's follow-through physics or reaction. Even Nvidia PhysX couldn't handle that with Tomb Raider hair and it slowed down the GPU pipeline considerably. That's one of the reasons we only see it with a demo of a single head with hair and never in a game.
 
Last edited:

Fafalada

Fafracer forever
That is beyond realtime hardware.
With traditional approaches yes - but this potentially falls into use-cases CamHostage CamHostage alludes to above. For most part hair simulation can be treated as its own subspace, which could fit an ML model that would handle high-fidelity localized collisions, and interactions with external influences would be done with an actual physics solver.
 

VFXVeteran

Banned
With traditional approaches yes - but this potentially falls into use-cases CamHostage CamHostage alludes to above. For most part hair simulation can be treated as its own subspace, which could fit an ML model that would handle high-fidelity localized collisions, and interactions with external influences would be done with an actual physics solver.
When it happens, I'll believe it. Realtime graphics has always shown demos of advanced tech never to be seen in practice. When ms has to be divided it's just too much of a resource hog and bandwidth drain to be practical.
 
Last edited:

Kuranghi

Gold Member
That is beyond realtime hardware. You are talking about completely bounding every single strand of hair (or hair clump) and testing for collisions. The more accurate you want it to be the more intersection tests with small bounds will be needed. After that what do you do when you have to compute it's follow-through physics or reaction. Even Nvidia PhysX couldn't handle that with Tomb Raider hair and it slowed down the GPU pipeline considerably. That's one of the reasons we only see it with a demo of a single head with hair and never in a game.

I appreciate your input but I wasn't talking about anything you mentioned. I didn't say it needed a fully accurate simulation to fix the problem, and I don't think it does. I'm sure there is some in between of what we have now and a fully accurate simulation.
 

VFXVeteran

Banned
I appreciate your input but I wasn't talking about anything you mentioned. I didn't say it needed a fully accurate simulation to fix the problem, and I don't think it does. I'm sure there is some in between of what we have now and a fully accurate simulation.
Fixing the problem requires what I mentioned though - mainly the bounding box intersection tests.
 
Last edited:

Fafalada

Fafracer forever
When it happens, I'll believe it. Realtime graphics has always shown demos of advanced tech never to be seen in practice. When ms has to be divided it's just too much of a resource hog and bandwidth drain to be practical.
There's no denying that 'practicality' of things is often limited by performance, but it's not as black & white. Industry is littered with corpses of impractical approaches that performed perfectly fine, or even superior to surviving alternatives - but they were just not practical for a host of other reasons.
Aforementioned ML approaches to subspace simulation demonstrably perform well (as far as runtime costs go) but cost of creating them is 'very' steep right now, particularly because of how fiddly it all is to work with.
Time will tell if those costs ever come down enough to be considered production worthy - and 'then' it'll be a case to seeing if runtime benefits (if any, by that point) are still wortwhile.
 
Top Bottom